Sparse Autoencoders Enhance Interpretable Out-of-Distribution Detection

Reliable detection of out-of-distribution (OOD) samples is crucial for safely deploying machine learning models, as neural networks often make overconfident predictions on unfamiliar inputs. While many OOD detection methods focus on the final output, they often overlook the rich, hierarchical information within intermediate network layers. This paper introduces a novel post-hoc detection approach that employs sparse autoencoders (SAEs) to extract interpretable features from these intermediate activations. The key insight is that in-distribution (ID) and OOD data activate distinct sets of these sparse features. A new OOD score, derived from the cosine similarity between a test sample's sparse feature activations and the mean activations of ID classes, is proposed. This method not only achieves state-of-the-art performance on standard OOD benchmarks but also offers valuable, interpretable insights into how distribution shifts manifest in learned representations.